
Picture a writer who has used AI for a year. The tool has been right often enough that checking its drafts starts to feel like a formality. The team lead finds out later, when an error reaches a reader.
The study
Researchers from Carnegie Mellon University and Microsoft Research surveyed 319 knowledge workers, recruited online through Prolific, who use generative AI at work at least once a week. Participants described 936 real examples of using it in their tasks, and for each one reported whether they thought critically and how much effort that took (Lee et al., The Impact of Generative AI on Critical Thinking, CHI 2025).
Participants reported thinking critically in 555 of the 936 examples, or 59% (Lee et al.). In their examples, critical thinking included setting clear goals, refining prompts and assessing AI output against criteria.
Two findings matter most for a team:
Confidence in the tool went with less critical thinking. The authors write that "higher confidence in GenAI is associated with less critical thinking" (Lee et al.).
Confidence in one's own skills went with more. Workers confident in their own abilities reported more critical thinking, and reported that it took more effort, particularly when evaluating and applying AI responses (Lee et al.).
The study also describes how the work changes. In the authors' words, generative AI "shifts the nature of critical thinking toward information verification, response integration, and task stewardship," as knowledge workers move from carrying out tasks to overseeing them (Lee et al.).
The paper also lists what pushed participants to think critically: wanting better work, avoiding negative outcomes and developing their skills. It lists what held them back: not realising a check was needed, time pressure or a sense that checking was outside their job, and difficulty improving AI responses in an unfamiliar subject (Lee et al.).
All of this is self-reported, from an online sample. The study measures what people say they do, and it shows associations, so it cannot tell us that confidence in AI causes less critical thinking.
What it means for a team adopting AI
We read verification, integration and stewardship as a description of the job a writer, instructional designer or editor now does alongside drafting.
Verification: checking facts, figures and claims in the AI output against sources the team trusts.
Integration: fitting the output into the document, the template, the voice and the rest of the content set.
Stewardship: owning the task from the brief to sign-off, including deciding what the AI should and should not do. Our post on writing the spec before you ask AI to build anything covers one part of this.
These are the steps we suggest.
Name these activities in roles and reviews. Time pressure was one of the barriers participants reported (Lee et al.). An expectation that is written into the role is easier to protect when deadlines tighten.
Build subject knowledge on purpose. Confidence in one's own abilities went with more critical thinking, and difficulty in an unfamiliar subject was one of the barriers (Lee et al.). Time spent learning the product and the audience supports better checking.
Decide which tasks always get a full check. The authors note that AI may reduce critical engagement, particularly in routine or lower-stakes tasks (Lee et al.). Agree as a team which content is never treated as routine: regulated statements, figures, procedures and anything customer-facing.
Budget time for checking. If AI saves drafting time and all of it goes into more output, the checking time disappears.
Reliable AI Workflows (Programme 02) includes a module on AI output evaluation and review, and uses your team's own documents, tools and review processes. To discuss your team, book a discovery call.


